That is a great topic.
We do need NN and machine learning applications written by modern Fortran as more as possible, that is without question.
If someone could show one small minimal working example about NN that modern Fortran could beat Pytorch in terms of speed. It could be great.
Some very good small NN examples to begin with, could be found in Andrew Ag’s Coursera course such as
https://www.coursera.org/specializations/deep-learning?
Foe example, training a NN to recognize a cat, etc. Those exercises written by Python iare small enough and can be written in modern Fortran.
Furthermore, to go to NN, we may also need to learn/know how to use Fortran to fully utilize power of GPU.
On the other hand, from business point of view, in order to making Fortran continue to survive and thrive, one have to let more and more people use Fortran directly or indirectly, partially or fully, as easy as possible. I do agree with @certik at thread #4
To begin with making Fortran more easily to be called from other languages such as Python/Julia, or call them within Fortran, is a very important/urgent thing.